Check the files your coding agent obeys before it obeys them.

Coding agents load CLAUDE.md, AGENTS.md, .cursorrules and their config files on every run and follow them. A poisoned rules file from a template, a cloned repo or a pull request can hide instructions in invisible Unicode or an HTML comment: add a remote script to every page, send your keys somewhere, skip permission prompts, and never mention it. The scanner reads those files the way the agent does, ignores ordinary instructions, and flags only what should never be there, with the line and the decoded hidden text.

What you get back

• verdict (clean, review, dangerous) and a 0-100 risk score, per file and overall
• hidden text: Unicode tag characters (the rules-file backdoor), zero-width runs, bidi overrides and instructions inside HTML comments, decoded so you can read them
• risky instructions: curl | sh and encoded commands, --dangerously-skip-permissions and --no-verify, reading ~/.ssh, .env or wallets, planting